The Google Discover fan-out that surfaces niche sites

“Tailor Your Feed” is the primary time a person can form their Uncover feed by typing, in pure language, what they need to see. We now have tracked it from its first look in Search Labs^search-labs to the pipeline[^pipeline] that powers it. Ten key factors:

1. An explicit-control layer. Your immediate is changed into SEE_MORE / SEE_LESS actions, utilized after a feed refresh.
2. Seemingly an LLM[^llm] beneath the hood. A persistent chat thread, and your immediate changed into directions utilized to your feed (in actual time and over time).
3. The rebrand. “Tailor Your Feed” grew to become “Add subjects to your feed” in spring 2026, with a chat-style entry level.
4. The back-end pipeline. historicalnaturallanguagetuningcontent.f[^pipeline-id], the “historic” twin of naturallanguagetuningcontent.f.
5. Two methods content material is chosen. Entity[^entity] / curiosity growth (the bulk) vs a query-intent[^query-intent] fan-out[^fan-out] (the minority), the latter being the GEO[^geo] mechanism inside Uncover.
6. Seen attribution. The “You requested to see” label, the “ensuing from pure language tuning” tag, and a immediate historical past in My Exercise.
7. Area of interest websites and small creators surfaced. Vegan recipe creators, Mississippi In the present day, a LinkedIn publish, area of interest Japanese-property blogs and, as an illustration of the retrieval’s[^retrieval] behaviour, publishers outdoors the same old mainstream (VentureBeat surfaced on a “area of interest websites” immediate, although not itself a small website).
8. A recognition bypass. This pipeline largely carries content material that had barely circulated in Uncover earlier than, the other of the basic pipelines that re-serve already-popular articles.
9. What it adjustments for publishers. Choice energy shifts to the person, opening a 3rd path to visibility for small, area of interest websites.
10. Nonetheless EN-only, nonetheless nascent. Search Labs US solely (FR ≈ 0%), adoption nonetheless early. What’s subsequent.

---

Methodology

This text combines two commentary streams:

• Discipline monitoring of the function within the Google app since December 2025: UI states, server responses, attribution tags, and feed behaviour after every “Refresh / Replace your feed”. Captured on our take a look at gadgets, US (English) Search Labs accounts.
• A detailed studying of the feed itself: every card might be traced again to the pipeline that chosen it. By isolating the playing cards served by historicalnaturallanguagetuningcontent.f, we describe how this pipeline behaves relative to the remainder of the feed, drawing on 1492.imaginative and prescient monitoring information.

Three deliberate notes on how we phrase issues:

• We describe distribution outcomes[^distribution], whether or not an article had ever circulated in Uncover earlier than, not uncooked viewers numbers. Once we say a card has “no prior Uncover distribution”, we imply we discover no hint of earlier serving in our Uncover monitoring dataset.
• No account identifier seems on this article. Examples are proven as immediate → consequence, anonymised.

The interior mechanisms under are our interpretation of noticed information and public analysis. The place a date is inferred relatively than anchored, we are saying so.

---

1. What “Tailor Your Feed” is: an explicit-control layer

For years, Uncover personalization was implicit: Google inferred your pursuits from clicks, dwell time, follows. “Tailor Your Feed” provides the other, an express layer the place you merely sort what you need.

The present entry level on the high of the feed: “What do you need to see?” with an “Add subjects to your feed” subject.

The function opens a chat-style panel. You’ll be able to choose a urged template or write freely.

The unique “Tailor your feed” intro card: “Say what you need in your personal phrases”, with a “Attempt now” button.

Instructed prompts: “Begin displaying me ladies’s basketball”, “Maintain me up to date on nation music”, “Present me extra of Cara Nicole’s movies”, with a free-text field “Ask for the sort of content material you need.”

These strategies correspond to 4 intents the server responses expose: SEE_MORE, KEEP_UPDATED, CREATOR_MORE and SEE_LESS. No matter you sort is interpreted into one (or a number of) of those actions, then utilized to the feed as soon as confirmed with “Refresh / Replace your feed.”

The function shipped by means of Search Labs, US solely.

The Search Labs entry: “Make your Uncover feed actually yours by saying what you need to see.” Beta, US solely.

Later iterations turned the free-text field into express starter templates: “Present me content material from…”, “I would like movies about…”, “Maintain me up to date…”. The identical verbs, now surfaced as chips.

The “What do you need to see?” panel with its starter templates: “Present me content material from…”, “I would like movies about…”, “Maintain me up to date…”.

---

2. Below the hood: seemingly an LLM that turns prompts into actions

The circulation is straightforward: immediate → interpretation → readable reply + an actionable consequence. You sort a immediate, the assistant replies in plain language and proposes concrete adjustments, and a faucet on “Replace your feed” commits them.

A consultant response, noticed within the information exchanges for the immediate “present me extra content material on seroundtable.com”, seems to be like this:

{
  "function": "Uncover • Tailor your feed",
  "locale": "en-US",
  "thread_key": "chat_thread_key_082fa565-234a-451c-9318-1e9af8b9d734",
  "user_prompt": "present me extra content material on seroundtable.com",
  "assistant_text": "I can present you extra content material from Search Engine Roundtable (SERoundtable.com) associated to your pursuits. Refresh your feed to use these adjustments.",
  "consequence": {
    "standing": "UNDERSTOOD_AND_ACTIONABLE",
    "actions": ["SEE_MORE"],
    "show_call_to_action": true,
    "rely": 1
  },
  "ui_state_code": 2
}

The in-app immediate that produced the response above: “present me extra content material on seroundtable.com”, and the assistant’s reply ending on “Refresh your feed.”

Three issues stand out:

• A persistent thread. The thread_key is secure throughout exchanges; your tuning is a dialog, not a one-shot. The identical thread is referenced once more when, later, a card is attributed to considered one of your previous prompts.
• Actions, not subjects. The response returns actions: ["SEE_MORE"]. Ask to take away a subject and also you get ["SEE_LESS"]; a nuanced immediate can return each, e.g. “new nation music releases… however no movie star gossip” yields ["SEE_MORE", "SEE_LESS"].
• Native context injection. Responses are interpreted together with your context (locale, language, location). A generic “maintain me up to date about NBA” got here again with “Updates on the Brooklyn Nets“, an area workforce injected from context.

“maintain me up to date about nba” → the assistant proposes scores, workforce and participant information, and “Updates on the Brooklyn Nets”, a locally-injected entity.

On Google’s aspect, your sentence is changed into a set of directions that feed the retrieval stage, with an “offline” path (utilized over time) and a real-time one. That is, as we learn it, the architectural shift the function embodies: from inferred curiosity vectors to a natural-language profile you write your self.

---

3. A six-month timeline (December 2025 → June 2026)

We documented the rollout because it occurred. Dates anchored to in-app captures and the function’s personal changelogs; a couple of are approximate (marked with “~”).

December 2025: first sighting (Search Labs, US). The function seems with every little thing described above: the chat panel, the JSON response, the persistent thread, the 4 intents, native context injection. First impression: the impact is delicate; after a number of refreshes you see a couple of on-topic playing cards, however nothing dramatic. (field note, example)

“I would like a break from destructive information. Present me extra feel-good tales, however maintain the native and breaking information.” The assistant lists the sorts of uplifting content material to count on, then provides “Refresh your feed.”

One other early instance: “are you able to present me https://dev.to/ on my feed” → the assistant provides programming articles, web-dev tutorials and software program information.

~January 13, 2026: the attribution tag. Google begins marking playing cards “ensuing from pure language tuning” after a refresh and the SEE_MORE/SEE_LESS arbitration, making it potential, for the primary time, to inform which playing cards a immediate truly modified. A immediate historical past additionally seems in My Exercise. (field note)

“‘Tailor your feed’ preferences in Uncover. View and handle your prompts.” A devoted My Exercise floor (right here, empty).

~February 2026: “historic” tuning, and what SEE_LESS actually does. A second tag seems, “historic pure language tuning”, for playing cards influenced by a previous immediate. Testing “fewer X posts”, we noticed Google change X (Twitter) playing cards with YouTube movies, and, notably, asking to take away a subject doesn’t actually take away it: you get SEE_LESS, however the subject isn’t deleted from the feed. (field note)

~February 2026: the “area of interest” take a look at. Requested for “extra area of interest / small websites”, the feed got here again, on the primary refresh, with 2 of 10 playing cards modified by the immediate (a one-off snapshot, not a median), surfacing VentureBeat and Mississippi In the present day, with the very first consequence pushed by the request. (field note)

~February 2026: the “entity” take a look at. Requested for “extra articles from a particular creator”, Google understood the subjects associated to that creator (entities), refreshed, and surfaced a LinkedIn publish from them, tagged “pure language tuning content material”. (field note)

~April 2026: the rebrand + the “You requested to see” label. “Tailor your feed” turns into “Add subjects to your feed” with a chat UI, and a visual “You requested to see” label now marks the playing cards served by the pipeline. (rebrand, label)

A card labeled “You requested to see”: a historical-figures listicle from AOL.

“You requested to see” on a Guardian politics story.

“You requested to see” on Search Engine Roundtable playing cards; the writer requested earlier now surfaces explicitly.

Might 22, 2026: the question intents (the fan-out). We affirm that, past entity growth, some pipeline playing cards carry a saved question intent, the immediate decomposed into particular retrieval queries, on the identical sample because the GEO “fan-out”. (Extra in part 5.) (field note)

“You requested to see” additionally applies to video: a Crunchyroll/One Piece YouTube card surfaced by a immediate.

June 2026: present state. The entry level reads “Add subjects to your feed”, and the function now surfaces small creators nicely outdoors the key publishers (part 6).

---

4. The pipeline behind it: historicalnaturallanguagetuningcontent.f

Each Uncover card might be traced again to the pipeline that chosen it. “Tailor Your Feed” maps to a devoted pair:

• naturallanguagetuningcontent.f, content material primarily based in your present natural-language preferences.
• historicalnaturallanguagetuningcontent.f, content material primarily based on previous prompts that maintain influencing the feed (the “historic” tag from the timeline).

The pipeline retrieves content material in two distinct methods:

• Mode A: entity / curiosity growth (the bulk). Primarily based on the noticed behaviour, the immediate is mapped to entities and subjects, and the feed expands round them. For this reason asking for one writer surfaces associated sources and subjects, not simply that writer, the identical logic because the Comply with button. Most playing cards work this fashion, increasing round your subjects relatively than echoing the precise phrases you typed.
• Mode B: query-intent fan-out (the minority). For a fraction of playing cards, the immediate is decomposed into express question intents, natural-language retrieval queries that fetch the article. That is the GEO “fan-out” mechanism, and it’s the topic of part 5.

One behaviour price flagging: in our monitoring information, Google seems to advertise these playing cards cautiously, on common lower than different pipelines, and pulls them again extra typically than another, in step with a retrieval that generally matches loosely (we’ll see a concrete false optimistic in part 5). It’s, by design, a focused pipeline, not a mass-distribution one: its playing cards present basically no development over time, the bottom of any pipeline. It serves what was requested for, to the person who requested. It doesn’t snowball.

---

5. Question intent: the GEO “fan-out” inside Uncover

That is essentially the most attention-grabbing mechanism. For a slice of playing cards, the immediate is damaged down into a particular question intent that matched the article, the immediate changed into exact, natural-language retrieval queries. It’s the purposeful analogue of the fan-out described for Generative Engine Optimization: a single immediate is decomposed into sub-queries that retrieve content material by semantic relevance, with no recognition prerequisite.

The decomposition is seen. A immediate about search engine optimisation turns into a set of informational queries:

Consumer immediate (approx.)	Decomposed question intents
“Present me content material from search engine optimisation”	“search engine optimisation methods algorithm adjustments” · “Google rating system updates” · “ideas for getting content material into google uncover”

And people question intents then retrieve actual articles. Listed here are anonymised question intent → URL pairs we noticed (the formulations are the precise inside question intents):

Question intent	Retrieved article	Profile
beginning seeds indoors information	buzzyseeds.com/…/how-to-grow-strawberry-from-seed-indoors	area of interest gardening, no prior Uncover distribution
shopping for Japanese property information	japantoday.com/…/how-to-buy-a-home-in-japan-as-a-foreigner	area of interest
shopping for Japanese property information	maigomika.com/…/rural-japan-inaka-levels	area of interest
private tales residing in France	perfectlyprovence.co/…	area of interest
ideas for getting content material into google uncover	conductor.com/academy/best-aeo-geo-tools	mid-size
AI content material machine studying search engine optimisation	searchengineland.com/ai-increase-seo-expertise-value	commerce press
greatest sci-fi books	collider.com/best-sci-fi-books-last-25-years-ranked	mainstream
Nvidia inventory evaluation	reuters.com/expertise/nvidia-invests-2-billion	mainstream, already extensively distributed, merely re-surfaced

Our analyzer teams these question intents into clusters:

Clusters of question intents noticed within the pipeline: “studying Google algorithms”, “AI content material machine studying search engine optimisation”, “shopping for Japanese property information”, “wholesome cooking strategies”, “anime suggestions 2026″…

The identical clusters, with the article rely per intent.

Drilling into one cluster reveals each the energy and the restrict of the fan-out:

The “shopping for Japanese property information” cluster: japantoday.com (tips on how to purchase a house in Japan) and maigomika.com (rural Japan) are spot-on area of interest matches, however rockpapershotgun.com (Forza Horizon 6 in-game dwelling areas) is a false optimistic, a video-game article pulled in by floor phrase overlap. Free matches like this are why Google ranks this pipeline so cautiously (part 4).

Why this issues for search engine optimisation: the question intent reveals the actual vocabulary Google makes use of to map a immediate to your content material. These are natural-language informational queries, not uncooked key phrases. Aligning titles, H1s and intros with these formulations is the Uncover-side equal of optimizing for the AI fan-out.

---

6. Area of interest websites and small creators: the recognition bypass

Basic Uncover pipelines largely re-serve content material that’s already common, articles which have already circulated extensively and constructed engagement. “Tailor Your Feed” works in a different way: the playing cards we observe present a retrieval that reaches for semantically related content material no matter whether or not it ever circulated in Uncover earlier than.

1492.imaginative and prescient monitoring information backs this up. On historicalnaturallanguagetuningcontent.f, a majority of the playing cards level to articles with no detectable prior Uncover distribution in our dataset, content material that had by no means (or barely) been served within the feed earlier than. That is, by a large margin, the best share of any pipeline: the basic information pipelines present the other profile, the place virtually each card has an extended distribution historical past. A minority of the pipeline’s playing cards are the exception, mainstream articles (a Reuters story, as an example) already extensively distributed and easily re-surfaced right here.

The clearest illustration is a recipe immediate. Asking for vegan recipes surfaced, not the large meals publishers, however small unbiased creators:

“Present me content material from recipes vegan” → the assistant proposes plant-based weeknight dinners, vegan stews, high-protein tofu, vegan baking… then “Replace your feed.”

The consequence: a “Candy Potato Tacos” recipe from an unbiased creator, and “72 Vegan BBQ Recipes” from a small vegan weblog, every labelled “You requested to see”, with a ranking widget (“How would you fee this suggestion?”).

Throughout our monitoring, the identical sample recurs: a distinct segment gardening weblog (buzzyseeds.com) for a seed-starting immediate; Mississippi In the present day for a “area of interest websites” immediate (with VentureBeat surfaced on the identical immediate, as an illustration of the retrieval’s behaviour); a LinkedIn publish for a creator immediate; area of interest Japanese-property blogs (japantoday.com, maigomika.com) for a property immediate. Most will not be the same old high-volume Uncover winners.

The takeaway, rigorously put: the function surfaces articles that had barely circulated in Uncover earlier than. The retrieval seems to be pushed by relevance to the immediate, not by prior recognition: what works like a recognition filter within the basic pipelines is, right here, bypassed.

---

7. What this adjustments for publishers

That is the half that issues most trying ahead, and it’s a real shift in how Uncover visibility might be earned.

Choice energy strikes to the person. Within the basic feed, Google decides what you see from inferred alerts. Right here, the person writes the immediate (“present me extra of X”, “much less of Y”), and Google turns it into entities, pursuits and question intents that drive retrieval. Demand turns into express.

Consequence: area of interest websites can floor and not using a Uncover observe file. As a result of the retrieval seems to succeed in for relevance relatively than prior recognition, a small website might be served the second a person asks for its subject, even when it had by no means actually circulated in Uncover earlier than (part 6). That’s new.

It’s a 3rd path to visibility. Till now, a distinct segment website broke into Uncover solely two methods: by means of sturdy implicit affinity (Google infers, from repeated engagement, that you just love a subject, and a re-surface pipeline retains feeding you that area of interest website), or by means of an express comply with. “Tailor Your Feed” provides a 3rd, user-initiated path that relies on neither.

The concrete levers for publishers:

• Optimize for entities/subjects (the dominant Mode A). Be unambiguously about what customers will identify. A transparent topical focus → cleaner entity affiliation → you’re within the growth set when somebody asks to your topic.
• Optimize for query-intent vocabulary (the Mode B fan-out). Phrase titles, H1s and intros to match the natural-language informational queries a immediate decomposes into (the Uncover-side of GEO). Part 5 reveals the precise formulations Google makes use of.

What it’s not. Publishers needs to be clear-eyed:

• Not a mass channel. The pipeline reveals basically no development, and Google promotes its playing cards cautiously. It serves the person who requested; it doesn’t broadcast.
• Not publisher-triggerable. Solely the person can hearth it. You might be retrieval-ready, however you’ll be able to’t activate it to your personal website.
• Geographically and adoption-limited. It’s EN-only (Search Labs, US; ≈ 0% in French feeds), and adoption continues to be early; the My Exercise floor was empty in our checks. The long run impression relies on (a) a common rollout and (b) whether or not customers truly undertake prompt-based tuning at scale.

The strategic learn: if “Add subjects to your feed” graduates from Search Labs and customers embrace it, the demand-driven, popularity-agnostic retrieval it depends on is structurally beneficial to small, centered, well-described websites, the sort that basic, popularity-dominated pipelines hardly ever reward.

---

What’s subsequent?

A snapshot, not an endpoint. What we’re watching:

• The French (and EU) rollout. The function is EN-only. Relying on how unbiased it’s from AIO[^aio] and AI Mode[^ai-mode] options, it might attain France in the end.
• Adoption. A strong mechanism with no customers adjustments nothing. The empty My Exercise floor suggests prompt-based tuning continues to be a distinct segment behaviour. Mass adoption is the variable that decides whether or not this issues for publishers.
• The present / historic pair. naturallanguagetuningcontent.f (dwell) and historicalnaturallanguagetuningcontent.f (persistent) counsel tuning is supposed to final over time, a standing instruction, not a one-off.
• Convergence with generative retrieval. A nascent generativeretrieval.f pipeline, noticed just lately in our monitoring information, suggests LLM-driven retrieval might attain past this one function (to be confirmed).

The larger image: Uncover is shifting from noticed personalization (Google infers) towards declared personalization (you inform it), and the retrieval that serves declared intent doesn’t lock onto recognition. That’s the structural opening for area of interest publishers, if and provided that the function ships broadly and customers undertake it.

---

Notes

[^pipeline]: In Uncover, a pipeline is a variety circuit that chooses and serves the playing cards. Every card might be traced to the identifier of the pipeline that produced it; that’s what we exploit in our 1492.imaginative and prescient monitoring information.

[^pipeline-id]: The .f suffix in identifiers reminiscent of historicalnaturallanguagetuningcontent.f is an inside marker we observe within the metadata of Uncover playing cards; it distinguishes the choice circuits.

[^search-labs]: Google’s beta program that lets customers take a look at experimental options in Search and Uncover earlier than any wider rollout.

[^llm]: Massive Language Mannequin. We assume one right here, with out direct proof of the mannequin used, from the conversational behaviour and the structured responses noticed.

[^entity]: Within the sense of Google’s Information Graph: a named entity; an individual, subject, organisation or idea that’s recognized and linked to others, distinct from the on a regular basis sense of the phrase “entity” (an organization, a authorized construction).

[^fan-out]: A mechanism by which a single immediate is damaged into a number of retrieval sub-queries, every focusing on a special angle of the subject.

[^query-intent]: A natural-language informational question, derived from the person’s immediate, that was used to retrieve a particular article. Observable in our information as metadata connected to sure pipeline playing cards.

[^geo]: Generative Engine Optimization: the observe of optimizing content material to be seen and cited within the solutions of generative engines (AI Overviews, ChatGPT, and many others.), notably by way of question fan-out.

[^retrieval]: The stage the place Google fetches and selects the content material to show within the feed, from curiosity alerts, entities, or queries.

[^distribution]: On this article, the truth that an article had already circulated in Uncover, as noticed in our 1492.imaginative and prescient monitoring dataset, not an viewers determine nor a Search Console metric.

[^aio]: AI Overviews: generative solutions proven on the high of sure Google Search outcomes.

[^ai-mode]: AI Mode: Google Search’s conversational interface, distinct from AI Overviews.

---

Discipline monitoring: Google app, Search Labs (US/English) accounts, December 2025 – June 2026. Pipeline behaviour derived from shut commentary of the Uncover feed by way of 1492.imaginative and prescient monitoring information. “Distribution” right here means whether or not an article had already circulated in Uncover, as noticed in our monitoring dataset, not non-public viewers figures. The interior mechanisms are our interpretation of noticed information and public analysis; inferred dates are flagged as approximate.